Universal object recognition
Large scale instance recognition is provided that can take advantage of channel-wise pooling. A received query image is processed to extract a set of features that can be used to generate a set of region proposals. The proposed regions of image data are processed using a trained classifier to classify the regions as object or non-object regions. Extracted features for the object regions are processed using feature correlation against extracted features for a set of object images, each representing a classified object. Matching tensors generated from the comparison are processed using a spatial verification network to determine match scores for the various object images with respect to a specific object region. The match scores are used to determine which objects, or types of objects, are represented in the query image. Information or content associated with the matching objects can be provided as part of a response.
1. A computer-implemented method, comprising:
obtaining image data;
determining, from the image data, a set of proposed object regions;
analyzing, using a trained classifier, the proposed object regions to determine a subset of proposed object regions having a highest likelihood of including representations of at least one object, the proposed object regions being pooled using a region-of-interest pooling network;
determining, from a set of object images, a subset of object images having a similarity to the proposed object regions;
performing spatial validation using correlated features of the subset of proposed object regions and each of the subset of object images to generate a set of similarity scores; and
providing information for a matching object based at least in part upon a respective similarity score between a respective proposed object region and a respective object image.
2. The computer-implemented method of claim 1 , further comprising:
generating, through feature correlation of the object region and each of the subset of object images, a set of matching tensors; and
processing the matching tensors in the spatial validation.
3. The computer-implemented method of claim 1 , wherein the proposed regions are pooled before classifying using the trained classifier.
4. The computer-implemented method of claim 3 , further comprising:
generating proposals for the subset of proposed object regions using a region proposal network (RPN), wherein the region-of-interest pooling network is configured to identify a proposed region as an object region using selective search on the proposals.
5. The computer-implemented method of claim 1 , further comprising:
receiving, from a client device, a search request; and
obtaining the image data from the search request.
6. The computer-implemented method of claim 5 , further comprising:
providing the information for the matching object as at least one search result in response to the search request.
7. The computer-implemented method of claim 1 , further comprising:
determining, from the image data, pixel coordinates defining bounding boxes for the subset of proposed object regions; and
providing the pixel coordinates for purposes of analyzing the proposed object regions using the trained classifier.
8. The computer-implemented method of claim 1 , further comprising:
providing a training data set including a maximum ratio of object images relative to images without represented objects, a number of object images in the training data being lower than a number of images without represented objects.
9. The computer-implemented method of claim 1 , further comprising:
performing feature extraction on the image data using a trained convolutional neural network.
10. The computer-implemented method of claim 1 , wherein the spatial validation is configured to identify objects of arbitrary types.
11. A system, comprising:
at least one processor; and
memory including instructions that, when executed by the at least one processor, cause the system to:
obtain image data;
determine, from the image data, a set of proposed object regions;
analyze, using a trained classifier, the proposed object regions to determine a subset of proposed object regions having a highest likelihood of including representations of at least one object;
identify, in the image data, the subset of proposed object regions;
pool the subset of regions using a region-of-interest pooling network;
determine, from a set of object images and using a proximity search, a subset of object images having a similarity to the proposed object region;
perform spatial validation using correlated features of the subset of proposed object regions and each of the subset of object images to generate a set of similarity scores; and
provide information for a matching object based at least in part upon a respective similarity score between a respective proposed object region and a respective object image.
12. The system of claim 11 , wherein the instructions when executed further cause the system to:
generate, through feature correlation of the object region and each of the subset of object images, a set of matching tensors; and
process the matching tensors in the spatial validation.
13. The system of claim 11 , wherein the proposed regions are pooled using a region-of-interest pooling network before classifying using the trained classifier.
14. The system of claim 13 , wherein the instructions when executed further cause the system to:
generate proposals for the subset of proposed object regions using a region proposal network (RPN), wherein the region-of-interest pooling network is configured to identify a proposed region as an object region using selective search on the proposals.
15. The system of claim 11 , wherein the instructions when executed further cause the system to:
receive, from a client device, a search request;
obtain the image data from the search request; and
provide the information for the matching object as at least one search result in response to the search request.
16. A non-transitory computer-readable storage medium storing instructions, the instructions when executed by a processor causing the processor to:
obtain image data;
determine, from the image data, a set of proposed object regions;
analyze, using a trained classifier, the proposed object regions to determine a subset of proposed object regions having a highest likelihood of including representations of at least one object;
identify, in the image data, the subset of proposed object regions;
pool the subset of regions using a region-of-interest pooling network;
determine, from a set of object images and using a proximity search, a subset of object images having a similarity to the proposed object region;
perform spatial validation using correlated features of the subset of proposed object regions and each of the subset of object images to generate a set of similarity scores; and
provide information for a matching object based at least in part upon a respective similarity score between a respective proposed object region and a respective object image.
17. The non-transitory computer-readable storage medium of claim 16 , wherein the instructions when executed further cause the processor to:
generate, through feature correlation of the object region and each of the subset of object images, a set of matching tensors; and
process the matching tensors in the spatial validation.
18. The non-transitory computer-readable storage medium of claim 16 , wherein the proposed regions are pooled using a region-of-interest pooling network before classifying using the trained classifier.
19. The non-transitory computer-readable storage medium of claim 18 , wherein the instructions when executed further cause the processor to:
generate proposals for the subset of proposed object regions using a region proposal network (RPN), wherein the region-of-interest pooling network is configured to identify a proposed region as an object region using selective search on the proposals.
20. The non-transitory computer-readable storage medium of claim 16 , wherein the instructions when executed further cause the processor to:
receive, from a client device, a search request;
obtain the image data from the search request; and
provide the information for the matching object as at least one search result in response to the search request.